Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Contour Detection and Hierarchical Image Segmentation
β Some Experiments
CS395T β Visual Recognition
Presented by Elad Liebman
Overview
β’ Understanding the method better: β The importance of thresholding the ouput β Understanding the inputs better β Understanding the UCM
β’ Pushing for boundaries: β Types of difficult inputs β Difficult input examples
Warm-up: choosing the right threshold
π = 0.4
Warm-up: choosing the right threshold
π = 0.1
Warm-up: choosing the right threshold
π = 0.8
So what goes into the OWT?
β’ From Contours to Regions: An Empirical Evaluation. P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. CVPR 2009.
β’ Contour Detection and Hierarchical Image Segmentation. P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. PAMI 2011
The Features
β’ Difference in feature channels on the two halves of a disc of radius π and orientation π.
β’ Feature channels in our case: β Color gradients β Brightness gradients β Texture gradients
β’ Comparison between the two disc halves using π2 distance.
The basic signals
β’ Color/brightness channels based on Lab color space.
β’ Repeatedly generated at different scales (different π radii values - π
2,π, 2π).
β’ All in all we get 13 different inputs.
Illustration n. 1: Color gradient π (red-green scale)
π =π2
π = βπ4
π =
π2
maxπ
{πππ}
Illustration n. 2: Color gradient π, different π value
maxπ
{πππ}
π = 0
π =π4
π = βπ4
Illustration 3: Brightness gradient (light-dark scale)
maxπ
{πππ}
π =π2
π = βπ4
π =
π2
Illustration 4: Texton Map
Multiscale ππ and Spectral/Global ππ β’ The features are linearly combined to create a
unified signal, the multiscale Pb input. β’ They are then βglobalizedβ spectrally. β’ An affinity matrix π€ is constructed with πππ
being the maximal value of πππ along the line connecting π and π.
β’ Generalized eigenvectors are then extracted and combined (after processing) with the local ππ data to create the final input for the OWT.
πππ vs. πππ
(In practice we sample 8 orientationsβ¦)
πππ = maxπ
{ }
π = [0,π8
,π4
,3π8
, β¦ ,7π8
]
(In practice we sample 8 orientationsβ¦)
πππ =
Playing with the orientations a bitβ¦
Playing with the orientations a bitβ¦
Using less than 8 orientations
2 orientations
Using less than 8 orientations
4 orientations (in this case itβs enough)
The Ultrametric Contour Map
β’ The basic idea is to organize and merge sub-regions hierarchically and iteratively.
β’ At each point we merge two regions with the weakest boundary between them.
β’ The result is a dendrogram of nested regions.
Example 1
Example 2
Potentially Difficult Input
β’ We would like to test the OWT-UCM method against problematic input.
β’ See where it might break.
warming up - blurring
Potentially Difficult Input II
β’ Blurring isnβt very interesting β doesnβt reflect the kind of challenges the algorithm is likely to face.
β’ What realistic problems exist? β Difficult patches β Complex, contour-rich images β Not enough color/texture information
Difficult Input β abstract art
π = 0.4
Difficult Input β abstract art
π = 0.1
Difficult(er) Input β abstract(er?) art
π = 0.4
Difficult(er) Input β abstract(er?) art
π = 0.1
Another Difficult Example β’ Occlusion β’ Similarity of objects in image:
β Similar textures β Similar colors β Contour lines blend
Another Difficult Example
π = 0.4
Another Difficult Example
π = 0.1
Another Difficult Example
π = 0.2
Sneaking a look under the hood
β’ Possibly we need to weight elements differently in certain casesβ¦
CGA1 CGB1
Textons
BG1
Summary β’ Studied some of the technical aspects of the
method: β Threshold selection β Understanding the input data β Illustration of UCM in action
β’ Tested the method against difficult input: β Problematic contours β Complex images β Cases where features arenβt informative enough β Many similar items occluding one another
References β’ From Contours to Regions: An Empirical Evaluation. P.
Arbelaez, M. Maire, C. Fowlkes, and J. Malik. CVPR 2009. β’ Contour Detection and Hierarchical Image Segmentation.
P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. PAMI 2011
β’ Using Contours to Detect and Localize Junctions in Natural Images. M. Maire, P. Arbelaez, C. Fowlkes, and J. Malik. CVPR 2008.
β’ What are Textons? S. Zhu, C. Guo, Y. Want Z. Xu, International Journal of Computer Vision 2005.
β’ Class notes for 16-721: Learning-Based Methods in Vision, taught by Prof. Alexei Efros, CMU.
β’ Class notes for CS 143: Introduction to Computer Vision, taught by Prof. James Hayes, Brown.